Method and device for training risk prediction model and performing risk prediction on user

Abstract

The embodiment of the invention provides a method and a device for training a risk prediction model and performing risk prediction on a user. The method comprises the following steps: constructing a first feature vector according to available feature values of a sample user for preset N feature items; and performing filling operation on the first feature vector by using a missing value filling module to obtain a second feature vector. And then, processing the second feature vector by using a risk prediction model to obtain a risk prediction result of the sample user. And then, according to a risk prediction result and a risk label of a sample user, updating parameters in a missing value filling module and a risk prediction model. The sample user has feature value missing for at least one feature item in the N feature items. The missing value filling module comprises learnable parameters for the N feature items, and the filling operation comprises the step that the learnable parameters corresponding to at least one feature item serve as corresponding feature value codes.

Description

Technical Field

[0001] One or more embodiments of this specification relate to the field of machine learning, and in particular, to methods and devices for training risk prediction models and performing risk prediction on users. Background Art

[0002] Tabular data is one of the most commonly used data types in modeling. Missing values are a common problem within this field. For example, in credit risk control scenarios, there is a large amount of fragmented information, with missing features reaching as high as 80%. Furthermore, the missing features often vary across samples. In this case, filling in missing values can alter the data distribution of the entire dataset, thus affecting the model's generalization performance.

[0003] Therefore, it is expected that there will be an improved missing value filling scheme to improve model performance. Summary of the Invention

[0004] One or more embodiments of this specification describe a learnable filling mechanism, which uses a missing value filling module to explicitly perform missing value filling operations, and the missing value filling module and the risk prediction model are optimized together with the training objectives, so that the learnable parameters of the feature value encoding used to fill the missing feature values in the missing value filling module can be updated in the direction of the optimization target, thereby improving the missing value filling effect.

[0005] According to a first aspect, a training method for a risk prediction model is provided, comprising: constructing a first feature vector based on available feature values of sample users for N preset feature items, wherein the sample user has a missing feature value for at least one of the N feature items; performing a filling operation on the first feature vector using a missing value filling module to obtain a second feature vector; the missing value filling module includes learnable parameters for the N feature items, and the filling operation includes encoding the learnable parameters corresponding to the at least one feature item as corresponding feature values; processing the second feature vector using a risk prediction model to obtain a risk prediction result for the sample user; and updating the missing value filling module and the parameters in the risk prediction model based on the risk prediction result and the risk label of the sample user.

[0006] According to one embodiment, the first feature vector is disposed with a null value at an element position corresponding to the at least one feature item; and the filling operation includes filling the null value position with a learnable parameter corresponding to the at least one feature item.

[0007] According to one embodiment, the first feature vector has a preset default value at the element position corresponding to the at least one feature item; the filling operation includes replacing the default value with a learnable parameter corresponding to the at least one feature item.

[0008] According to one embodiment, the N feature items are all of numerical type; the second feature vector is an N-dimensional vector, and each dimension corresponds to the feature value code of a feature item.

[0009] According to one embodiment, the N feature items include at least one categorical feature item, the corresponding feature value of which is encoded as a one-hot vector.

[0010] According to one embodiment, the obtainable characteristic value includes the first numerical value of the sample user for the first characteristic item; the construction of the first characteristic vector includes: obtaining M quantile intervals divided for the first characteristic item; determining the first coding value corresponding to the first numerical value based on a first ratio of the sequence number of the target quantile interval in which the first numerical value falls and the number of quantile intervals M, and a second ratio representing the relative position of the first numerical value within the target quantile interval; the first coding value forms a part of the first characteristic vector.

[0011] In a further embodiment, the first coding value is the sum of a third ratio obtained by dividing the second ratio by the number M and the first ratio.

[0012] According to one embodiment, the risk prediction model includes several target networks consisting of two adjacent processing layers; any second node in the output layer of the target network aggregates the activation values of the first nodes of the input layer for the second node, and the activation value is obtained by processing the characteristic value output by the first node through a learnable first activation function between the first node and the second node; the first activation function is obtained by combining the basic functions through combination parameters; the model parameters in the risk prediction model include the combination parameters.

[0013] Furthermore, the risk prediction model also includes multiple target processing layers, and the model processing also includes: the target processing layer obtains the output feature of this layer by calculating the product of the second activation value and the gating signal; wherein the second activation value is obtained by processing the first mapping value using a preset second activation function, the first mapping value is the result of a first linear transformation of the input feature of this layer, and the gating signal is the result of a second linear transformation of the input feature of this time.

[0014] According to a second aspect, a method for predicting risks for a user is provided, comprising: constructing a first feature vector based on the available feature values of the target user for N preset feature items, wherein the target user has a missing feature value for at least one of the N feature items; performing a filling operation on the first feature vector using a missing value filling module to obtain a second feature vector; the missing value filling module includes learnable parameters for the N feature items, and the filling operation includes encoding the learnable parameters corresponding to the at least one feature item as corresponding feature values; the learnable parameters are determined by training the risk prediction model and the missing value filling module; and processing the second feature vector using the risk prediction model to obtain a risk prediction result for the target user.

[0015] According to a third aspect, a training device for a risk prediction model is provided, comprising: a construction unit, configured to construct a first feature vector based on available feature values of a sample user for N preset feature items, wherein the sample user has a missing feature value for at least one of the N feature items; a filling unit, configured to perform a filling operation on the first feature vector using a missing value filling module to obtain a second feature vector; the missing value filling module includes learnable parameters for the N feature items, and the filling operation includes encoding the learnable parameters corresponding to the at least one feature item as corresponding feature values; a processing unit, configured to process the second feature vector using a risk prediction model to obtain a risk prediction result for the sample user; and an updating unit, configured to update the missing value filling module and the parameters in the risk prediction model based on the risk prediction result and the risk label of the sample user.

[0016] According to a fourth aspect, a device for predicting risks for users is provided, comprising: a construction unit, configured to construct a first feature vector based on available feature values of a target user for preset N feature items, wherein the target user has a missing feature value for at least one of the N feature items; a filling unit, configured to perform a filling operation on the first feature vector using a missing value filling module to obtain a second feature vector; the missing value filling module includes learnable parameters for the N feature items, and the filling operation includes encoding the learnable parameters corresponding to the at least one feature item as corresponding feature values; the learnable parameters are determined by training the risk prediction model and the missing value filling module; a processing unit, configured to process the second feature vector using the risk prediction model to obtain a risk prediction result for the target user.

[0017] According to a fifth aspect, a computer program product is provided, comprising a computer program / instruction, which implements the steps of the method according to the first aspect or the second aspect when executed by a processor.

[0018] According to a sixth aspect, a computer-readable storage medium is provided, on which a computer program is stored. When the computer program is executed in a computer, the computer is caused to execute the method described in the first aspect or the second aspect.

[0019] According to the seventh aspect, a computing device is provided, comprising a memory and a processor, wherein executable code is stored in the memory, and when the processor executes the executable code, the method of the first aspect or the second aspect is implemented.

[0020] In an embodiment of the present specification, when a sample user has a missing eigenvalue for at least one of the N eigenvalues, a missing value filling module is used to fill in the first eigenvector to obtain a second eigenvector. Furthermore, the missing value filling module is optimized along with the risk prediction model along with the training objective. Therefore, when a sample contains a missing eigenvalue, the filling value corresponding to the eigenvalue encoding of the missing eigenvalue in the missing value filling module is updated toward the optimization objective. In this way, through training, a suitable filling value for the eigenvalue encoding of each missing eigenvalue can be obtained. BRIEF DESCRIPTION OF THE DRAWINGS

[0021] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the following briefly introduces the drawings required for use in the description of the embodiments. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0022] Figure 1 A schematic diagram of the technical concept of a learnable filling mechanism is shown.

[0023] Figure 2 A flowchart of a method for training a risk prediction model according to one embodiment is shown.

[0024] Figure 3 A flow chart of a method for predicting risk for a user according to one embodiment is shown.

[0025] Figure 4 A schematic structural diagram of a risk prediction model training device according to one embodiment is shown.

[0026] Figure 5 A schematic structural diagram of an apparatus for predicting risk for a user according to one embodiment is shown. DETAILED DESCRIPTION

[0027] The solution provided in this specification is described below in conjunction with the accompanying drawings.

[0028] Throughout this specification, words such as "exemplary," "for example," or "for instance" are used to indicate examples, illustrations, or descriptions. Any embodiment or design described in this specification as "exemplary," "for example," or "for instance" should not be construed as being preferred or advantageous over other embodiments or designs. Rather, the use of words such as "exemplary," "for example," or "for instance" is intended to present the relevant concepts in a concrete manner.

[0029] In this specification, the term "and / or" is simply a description of an association between related objects, indicating that three possible relationships exist. For example, "A and / or B" can mean: A exists alone, B exists alone, and A and B exist simultaneously. In addition, unless otherwise specified, the term "plurality" means two or more.

[0030] Furthermore, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly identifying the technical features being referred to. Thus, features specified as "first" or "second" may explicitly or implicitly include one or more of such features. The terms "include," "comprising," "having," and their variations all mean "including but not limited to," unless otherwise specifically emphasized.

[0031] Missing value filling schemes can be divided into constant filling and model-based filling.

[0032] Filling missing values with constant values is a common filling scheme. For example, missing values can be filled with feature statistics (such as the median or mean), or with values outside the distribution (i.e., unrelated to the feature distribution) to significantly distinguish missing values from normal values. Constant filling schemes have almost no additional computational cost, but they are not suitable for all cases, and sometimes inappropriate filling can severely disrupt the original distribution of features, resulting in poor generalization of the trained model.

[0033] Another direction for filling missing values is model-based filling. Model-based filling includes methods such as predicting or generating missing values through models such as KNN (K-Nearest Neighbor), regression models, and GAN (Generative Adversarial Network). For example, similar data points are found through KNN and their values are used to interpolate the filling values; missing values are predicted through tree models; and filling values are generated through generative models such as GAN and Diffusion Model. Compared with constant filling, model-based filling usually has better results. However, these models are difficult to train and have high computational costs. Most importantly, since the training of these models is irrelevant to the final task, it is difficult to effectively evaluate their filling effects.

[0034] In view of this, the embodiments of this specification propose a learnable imputation (L-Imp) mechanism. Figure 1 A schematic diagram showing the technical concept of a learnable filling mechanism.

[0035] like Figure 1 As shown, a missing value filling module is added to the risk prediction model, and the missing value filling module is used to explicitly perform the missing value filling operation.

[0036] For example, if the complete model input contains N feature items, the missing value filling module can include N learnable parameters, each corresponding to a feature item. A learnable parameter is used as a filler for the feature value encoding of the missing feature item when the feature value of the corresponding feature item is missing. In other words, missing feature value encodings in the model input built for a sample user can be replaced with the corresponding learnable parameter.

[0037] During the training phase, the missing value filling module and the risk prediction model are optimized together with the training objectives. Specifically, after the risk prediction model processes the model input filled with the missing value filling module and obtains the risk prediction result, the risk prediction model and the missing value filling module can be updated based on the risk prediction result and the true risk label.

[0038] Therefore, when there are missing features in the sample, the filling value corresponding to the missing feature value encoding in the missing value filling module will be updated in the direction of the optimization target. In this way, through training, the appropriate filling value for the feature value encoding of each missing feature value can be obtained.

[0039] The learnable padding mechanism proposed in the embodiments of this specification can be fully integrated into end-to-end training, with minimal increase in computational effort, especially during the inference phase, which is crucial in industrial applications. Specifically, the learnable padding mechanism can be merged with the model inference phase during deployment, requiring no additional model deployment and requiring minimal computational effort.

[0040] The specific implementation process of the above technical concept is described below.

[0041] Figure 2 A flow chart of a method for training a risk prediction model according to one embodiment is shown. It should be understood that the method can be executed by any device, equipment, platform, or device cluster with computing and processing capabilities.

[0042] See also Figure 2 In step S21, a first feature vector is constructed based on the available feature values of the sample user for the preset N feature items, wherein the sample user has a missing feature value for at least one of the N feature items.

[0043] Sample users refer to users whose true risk labels are known.

[0044] The preset N feature items (such as attribute items) can refer to a set of feature items that contribute to improving model performance through feature selection. Therefore, each of the N feature items can be regarded as an input feature.

[0045] The missing feature items in all samples used in the training phase may cover N feature items, so that the filling value (ie, the learnable parameter) corresponding to each feature item in the N feature items may be learned in the training phase.

[0046] The eigenvalue code of each obtainable eigenvalue may form a part of the first eigenvector. For example, the eigenvalue codes of all obtainable eigenvalues may be concatenated together to obtain the first eigenvector.

[0047] In some embodiments, the first feature vector may further identify feature items with missing feature values. For example, a special value (such as a null value or a preset default value) may be used in the first feature vector to represent feature items with missing feature values.

[0048] Thus, each feature item may have a corresponding element position in the first feature vector. Exemplarily, the number of elements corresponding to each feature item in the first feature vector may be equal to the length of the feature value code of the feature item.

[0049] In some embodiments, the N feature items are all of numerical type. The feature value encoding of the feature item of numerical type is still numerical. Taking the N feature items as an example, the first feature vector can be an N-dimensional vector. Each dimension (i.e., each element position) in the first feature vector can correspond to a feature item. The element value at the element position corresponding to the feature item whose feature value is not missing in the first feature vector is the encoding value of the obtainable feature value of the feature item. The element value at the element position corresponding to the feature item whose feature value is missing in the first feature vector can be a null value or a preset default value.

[0050] Data used in credit risk control scenarios mostly exists in tabular form, and most features in tabular data are numerical. Due to its heterogeneity, each column of a feature represents a different meaning. For example, tabular data can contain features with large numerical ranges (such as annual income) as well as features with smaller numerical ranges (such as the proportion of expenditures under different expenditure items). Therefore, it is necessary to find a suitable encoding method to encode numerical features.

[0051] In view of this, the embodiments of this specification also propose an encoding method suitable for features of numerical type.

[0052] The encoding principle of this encoding method can be summarized as follows: first, the values in each column are segmented, and then the values within the same segment (i.e., quantile interval) are mapped. In this way, the distribution of the original data is mapped to a uniform distribution, and the data within the same quantile interval can be effectively distinguished, and the influence of extreme values can be eliminated.

[0053] Taking the example of the available characteristic value including the first numerical value of the sample user for the first characteristic item, the N quantile intervals divided for the first characteristic item can be obtained first. That is, the data under the characteristic dimension of "first characteristic item" (such as all sample data) can be evenly divided into N segments using N-1 quantile points to obtain N quantile intervals. In this way, the distribution of the original data can be mapped to a uniform distribution. Then, the first coding value corresponding to the first numerical value can be determined based on the first ratio of the sequence number of the target quantile interval in which the first numerical value falls and the number of quantile intervals N, and the second ratio representing the relative position of the first numerical value in the target quantile interval. The first coding value forms part of the first eigenvector. For example, the element value at the element position corresponding to the first characteristic item in the first eigenvector is the first coding value.

[0054] In some embodiments, the first encoding value is the sum of the third ratio obtained by dividing the second ratio by the number N and the first ratio. For example, the encoding formula of the characteristic value (ie, the numerical value) of the characteristic item of the numerical type can be expressed as:

[0055]

[0056] Where x represents the eigenvalue of the numerical type (such as the first numerical value mentioned above); QLE(x) represents the coded value (such as the first coded value mentioned above); i represents the sequence number of the target quantile interval that x falls into; b i 、b i+1 Represents the two adjacent critical points (i.e., quantiles) corresponding to the i-th quantile interval.

[0057] In some embodiments, in addition to being numerical feature items (such as month, registration duration), feature items can also be categorical feature items (such as region). For example, N feature items may include at least one categorical feature item. For the feature values of categorical feature items, a one-hot encoding method can be used to obtain the corresponding encoding value. That is, the feature value encoding corresponding to the categorical feature item can be a one-hot vector. Exemplarily, the dimension of the one-hot vector can be L+1, where L is equal to the number of categories and 1 is a newly added bit, indicating whether there is a missing feature value for the categorical feature item.

[0058] In step S22, a missing value filling module is used to perform a filling operation on the first feature vector to obtain a second feature vector; the missing value filling module includes learnable parameters for N feature items, and the filling operation includes encoding the learnable parameters corresponding to at least one feature item as the corresponding feature value.

[0059] The first eigenvector is padded using the missing value filling module. Specifically, the learnable parameters corresponding to the eigenvalues of the missing eigenvalues of the eigenvalues are used as the eigenvalue codes of the missing eigenvalues of the eigenvalues, so that the second eigenvector obtained after padded contains the eigenvalue codes of all eigenvalues. For the eigenvalue codes of the available eigenvalues in the first eigenvector, the corresponding learnable parameters in the missing value filling module are not used for padded and remain unchanged.

[0060] In some embodiments, the first feature vector treats an element position corresponding to the at least one feature item as a null value; and the filling operation may include filling the null value position with a learnable parameter corresponding to the at least one feature item.

[0061] In some embodiments, the first feature vector has a preset default value at an element position corresponding to the at least one feature item; the filling operation may include replacing the default value with a learnable parameter corresponding to the at least one feature item.

[0062] Exemplarily, the dimension of the second feature vector may be equal to the sum of the lengths of the feature value codes of the N feature items. Taking the N feature items as an example, the second feature vector may be an N-dimensional vector, with each dimension corresponding to the feature value code of a feature item.

[0063] In step S23, the second eigenvector is processed using the risk prediction model to obtain a risk prediction result for the sample user.

[0064] The risk prediction model can adopt any machine learning model that can predict the risk of users.

[0065] The following is an example explanation of the structure of the risk prediction model.

[0066] In some embodiments, the risk prediction model may include several target networks consisting of two adjacent processing layers. Any second node in the output layer of the target network aggregates the activation values of the first nodes of the input layer for the second node, and the activation value is obtained by processing the eigenvalue output by the first node through the learnable first activation function between the first node and the second node. The first activation function is obtained by combining the basic functions through the combination parameters. The model parameters in the risk prediction model include combination parameters. Among them, the basic function may include but is not limited to spline functions. By setting a learnable first activation function between the first node and the second node, different features (i.e., the eigenvalues output by different first nodes) can be applied with different activation functions, thereby adapting to the distribution differences of different features.

[0067] In some embodiments, the risk prediction model may also include multiple target processing layers. Model processing also includes: the target processing layer obtains the output feature of this layer by calculating the product of the second activation value and the gating signal; wherein the second activation value is obtained by processing the first mapping value using a preset second activation function, the first mapping value is the result of a first linear transformation of the input feature of this layer, and the gating signal is the result of a second linear transformation of the input feature of this layer. The target processing layer that introduces the gating mechanism can expand the depth of the network and improve the expressive power, scalability and generalization of the model. Among them, the model parameters involved in the first linear transformation and the second linear transformation (such as feature weights and biases) can all be learnable.

[0068] In some embodiments, when the risk prediction model includes both a target network and a target processing layer, the target network can be placed in front and the target processing layer in the back. That is, the target network in the shallow layer (i.e., in the front) can be used to extract local features, and the target processing layer in the deep layer (i.e., in the back) can be used to improve the data processing capability and adaptability of the model.

[0069] In some embodiments, the risk prediction model may also include other model components. For example, the risk prediction model may also include, but is not limited to, a batch normalization layer (BatchNorm) and a dropout layer (Dropout). Exemplarily, the feature vector may first be processed by a batch normalization layer, then by several target networks, then by a dropout layer, and finally by multiple target processing layers. The output of the last target processing layer is the risk prediction result.

[0070] In step S24, the missing value filling module and the parameters in the risk prediction model are updated according to the risk prediction results and the risk labels of the sample users.

[0071] The risk label represents the true risk of the sample user. Based on the risk prediction results and risk labels, a loss function representing the predicted loss can be constructed. The parameters in the missing value filling module and the risk prediction model are updated based on the loss function.

[0072] Figure 3 A flow chart of a method for predicting risk for a user according to one embodiment is shown. It should be understood that the method can be executed by any device, equipment, platform, or device cluster with computing and processing capabilities.

[0073] See also Figure 3 In step S31, a first feature vector is constructed based on available feature values of the target user for N preset feature items, wherein the target user has a missing feature value for at least one of the N feature items.

[0074] Target users refer to users whose risk labels are unknown and for whom risk prediction is required.

[0075] It should be known that when the target user's feature values for each of the N feature items are available, the first feature vector constructed based on the available feature values already contains the feature value codes of all feature items, so there is no need to fill the first feature vector.

[0076] In some embodiments, the available feature values include a first numerical value of a target user for a first feature item; constructing a first feature vector includes: obtaining M quantile intervals divided for the first feature item; determining a first coded value corresponding to the first numerical value based on a first ratio of the sequence number of the target quantile interval in which the first numerical value falls to the number M of quantile intervals, and a second ratio representing the relative position of the first numerical value within the target quantile interval; the first coded value forms part of the first feature vector. Exemplarily, the first coded value is the sum of a third ratio obtained by dividing the second ratio by the number M and the first ratio.

[0077] In step S32, a missing value filling module is used to perform a filling operation on the first feature vector to obtain a second feature vector; the missing value filling module includes learnable parameters for the N feature items, and the filling operation includes encoding the learnable parameters corresponding to at least one feature item as corresponding feature values; the learnable parameters are determined by training the risk prediction model and the missing value filling module.

[0078] In some embodiments, the first feature vector is disposed with a null value at an element position corresponding to the at least one feature item; and the filling operation includes filling the null value position with a learnable parameter corresponding to the at least one feature item.

[0079] In some embodiments, the first feature vector has a preset default value at an element position corresponding to the at least one feature item; and the filling operation includes replacing the default value with a learnable parameter corresponding to the at least one feature item.

[0080] In some embodiments, the N feature items are all of numerical type; the second feature vector is an N-dimensional vector, and each dimension corresponds to the feature value code of a feature item.

[0081] In some embodiments, the N feature items include at least one categorical feature item, whose corresponding feature value is encoded as a one-hot vector.

[0082] In step S33, the second eigenvector is processed using the risk prediction model to obtain a risk prediction result for the target user.

[0083] In some embodiments, the risk prediction model includes several target networks consisting of two adjacent processing layers; any second node in the output layer of the target network aggregates the activation values of the first nodes of the input layer for the second node, and the activation value is obtained by processing the characteristic value output by the first node through a learnable first activation function between the first node and the second node; the first activation function is obtained by combining the basic functions through combination parameters; the model parameters in the risk prediction model include combination parameters.

[0084] In a further embodiment, the risk prediction model may also include multiple target processing layers, and the model processing also includes: the target processing layer obtains the output feature of this layer by calculating the product of the second activation value and the gating signal; wherein the second activation value is obtained by processing the first mapping value using a preset second activation function, the first mapping value is the result of a first linear transformation of the input feature of this layer, and the gating signal is the result of a second linear transformation of the input feature of this layer.

[0085] According to another embodiment, a training device for a risk prediction model is provided. Figure 4A schematic structural diagram of a risk prediction model training device according to one embodiment is shown. The device can be deployed in any device, platform or device cluster with data storage, computing and processing capabilities.

[0086] See also Figure 4 , a risk prediction model training device 400, comprising:

[0087] A construction unit 41 is configured to construct a first feature vector based on available feature values of a sample user for N preset feature items, wherein the sample user has a missing feature value for at least one of the N feature items;

[0088] a filling unit 42 configured to perform a filling operation on the first feature vector using a missing value filling module to obtain a second feature vector; the missing value filling module includes learnable parameters for the N feature items, and the filling operation includes encoding the learnable parameter corresponding to the at least one feature item as a corresponding feature value;

[0089] The processing unit 43 is configured to process the second feature vector using a risk prediction model to obtain a risk prediction result for the sample user;

[0090] The updating unit 44 is configured to update the missing value filling module and the parameters in the risk prediction model according to the risk prediction result and the risk label of the sample user.

[0091] In some embodiments, the first feature vector treats the element position corresponding to the at least one feature item as a null value; the filling unit 42 is specifically configured to fill the null value position with the learnable parameter corresponding to the at least one feature item.

[0092] In some embodiments, the first feature vector has a preset default value at the element position corresponding to the at least one feature item; the filling unit 42 is specifically configured to replace the default value with a learnable parameter corresponding to the at least one feature item.

[0093] In some embodiments, the N feature items are all of numerical type; the second feature vector is an N-dimensional vector, and each dimension corresponds to the feature value code of a feature item.

[0094] In some embodiments, the N feature items include at least one categorical feature item, the corresponding feature value of which is encoded as a one-hot vector.

[0095] In some embodiments, the obtainable characteristic value includes the first numerical value of the sample user for the first characteristic item; the construction unit 41 is specifically configured to: obtain M quantile intervals divided for the first characteristic item; determine the first coding value corresponding to the first numerical value based on the first ratio of the sequence number of the target quantile interval in which the first numerical value falls and the number of quantile intervals M, and the second ratio representing the relative position of the first numerical value in the target quantile interval; the first coding value forms a part of the first characteristic vector.

[0096] In a further embodiment, the first coding value is the sum of a third ratio obtained by dividing the second ratio by the number M and the first ratio.

[0097] For the specific implementation examples of each unit in the above device, please refer to the previous Figure 2 Description.

[0098] According to another embodiment, a device for predicting risk for a user is provided. Figure 5 A schematic structural diagram of an apparatus for predicting risk for a user according to one embodiment is shown. The apparatus can be deployed in any device, platform, or device cluster with data storage, computing, and processing capabilities.

[0099] See also Figure 5 , a device 400 for predicting risk for a user, comprising:

[0100] A constructing unit 51 is configured to construct a first feature vector based on available feature values of a target user for N preset feature items, wherein the target user has a missing feature value for at least one feature item among the N feature items;

[0101] a filling unit 52 configured to perform a filling operation on the first feature vector using a missing value filling module to obtain a second feature vector; the missing value filling module includes learnable parameters for the N feature items, and the filling operation includes encoding the learnable parameters corresponding to at least one feature item as corresponding feature values; the learnable parameters are determined by training the risk prediction model and the missing value filling module;

[0102] The processing unit 53 is configured to process the second feature vector using a risk prediction model to obtain a risk prediction result of the target user.

[0103] In some embodiments, the first feature vector treats the element position corresponding to the at least one feature item as a null value; the filling unit 52 is specifically configured to fill the null value position with the learnable parameter corresponding to the at least one feature item.

[0104] In some embodiments, the first feature vector has a preset default value at the element position corresponding to the at least one feature item; the filling unit 52 is specifically configured to replace the default value with a learnable parameter corresponding to the at least one feature item.

[0105] In some embodiments, the N feature items are all of numerical type; the second feature vector is an N-dimensional vector, and each dimension corresponds to the feature value code of a feature item.

[0106] In some embodiments, the N feature items include at least one categorical feature item, the corresponding feature value of which is encoded as a one-hot vector.

[0107] In some embodiments, the obtainable feature value includes the first numerical value of the target user for the first feature item; the construction unit 51 is specifically configured to: obtain M quantile intervals divided for the first feature item; determine the first coding value corresponding to the first numerical value based on the first ratio of the sequence number of the target quantile interval in which the first numerical value falls and the number of quantile intervals M, and the second ratio representing the relative position of the first numerical value in the target quantile interval; the first coding value forms a part of the first feature vector.

[0108] In a further embodiment, the first coding value is the sum of a third ratio obtained by dividing the second ratio by the number M and the first ratio.

[0109] For the specific implementation examples of each unit in the above device, please refer to the previous Figure 3 Description.

[0110] According to another embodiment, there is also provided a computer readable storage medium having a computer program stored thereon, which, when executed in a computer, causes the computer to execute a combination of Figure 2 or Figure 3 The method described.

[0111] According to another embodiment, a computer program product is also provided, including a computer program / instruction, which implements the aforementioned combination when executed by a processor. Figure 2 or Figure 3 The method steps described.

[0112] According to another embodiment, a computing device is provided, including a memory and a processor, wherein the memory stores executable code, and when the processor executes the executable code, the system realizes the combination of Figure 2 or Figure 3 The method described.

[0113] Those skilled in the art will appreciate that, in one or more of the above examples, the functions described herein may be implemented using hardware, software, firmware, or any combination thereof. When implemented using software, these functions may be stored in a computer-readable medium or transmitted as one or more instructions or codes on a computer-readable medium.

[0114] The specific implementation methods described above further illustrate the objectives, technical solutions and beneficial effects of the present invention in detail. It should be understood that the above description is only a specific implementation method of the present invention and is not intended to limit the scope of protection of the present invention. Any modifications, equivalent replacements, improvements, etc. made on the basis of the technical solution of the present invention should be included in the scope of protection of the present invention.

Claims

1. A method for training a risk prediction model, comprising: Constructing a first feature vector based on available feature values of the sample user for N preset feature items, wherein the sample user has a missing feature value for at least one of the N feature items; Performing a filling operation on the first feature vector using a missing value filling module to obtain a second feature vector; the missing value filling module includes learnable parameters for the N feature items, and the filling operation includes encoding the learnable parameter corresponding to at least one feature item as a corresponding feature value; Processing the second eigenvector using a risk prediction model to obtain a risk prediction result for the sample user; According to the risk prediction results and the risk labels of the sample users, the missing value filling module and the parameters in the risk prediction model are updated.

2. The method according to claim 1, wherein The first feature vector is disposed as a null value at the element position corresponding to the at least one feature item; The filling operation includes filling the empty value position with a learnable parameter corresponding to the at least one feature item.

3. The method according to claim 1, wherein The first feature vector has a preset default value at the element position corresponding to the at least one feature item; The filling operation includes replacing the default value with a learnable parameter corresponding to the at least one feature item.

4. The method according to claim 1, wherein The N feature items are all of numerical type; the second feature vector is an N-dimensional vector, and each dimension corresponds to the feature value code of a feature item.

5. The method according to claim 1, wherein The N feature items include at least one categorical feature item, whose corresponding feature value is encoded as a one-hot vector.

6. The method of claim 1, wherein The obtainable feature value includes a first value of the sample user for the first feature item; and constructing the first feature vector includes: Obtaining M quantile intervals divided for the first feature item; Based on a first ratio of the sequence number of the target quantile interval in which the first value falls to the number M of quantile intervals, and a second ratio representing the relative position of the first value within the target quantile interval, a first coding value corresponding to the first value is determined; the first coding value forms part of the first feature vector.

7. The method according to claim 6, wherein: The first coding value is the sum of a third ratio obtained by dividing the second ratio by the number M and the first ratio.

8. The method of claim 1, wherein The risk prediction model includes a plurality of target networks consisting of two adjacent processing layers; any second node in the output layer of the target network aggregates the activation values of the first nodes of the input layer for the second node, and the activation values are obtained by processing the feature values output by the first nodes through a learnable first activation function between the first nodes and the second nodes; The first activation function is obtained by combining basic functions through combination parameters; the model parameters in the risk prediction model include the combination parameters.

9. The method according to claim 8, wherein The risk prediction model also includes multiple target processing layers, and the model processing also includes: the target processing layer obtains the output feature of this layer by calculating the product of the second activation value and the gating signal; wherein the second activation value is obtained by processing the first mapping value using a preset second activation function, the first mapping value is the result of a first linear transformation of the input feature of this layer, and the gating signal is the result of a second linear transformation of the input feature of this time.

10. A method for predicting risk for a user, comprising: Constructing a first feature vector based on available feature values of a target user for N preset feature items, wherein the target user has a missing feature value for at least one of the N feature items; Performing a filling operation on the first feature vector using a missing value filling module to obtain a second feature vector; the missing value filling module includes learnable parameters for the N feature items, and the filling operation includes encoding the learnable parameter corresponding to at least one feature item as a corresponding feature value; The learnable parameters are determined by training the risk prediction model and the missing value filling module; The second eigenvector is processed using a risk prediction model to obtain a risk prediction result of the target user.

11. A training device for a risk prediction model, comprising: a constructing unit configured to construct a first feature vector based on available feature values of a sample user for N preset feature items, wherein the sample user has a missing feature value for at least one feature item among the N feature items; a filling unit configured to perform a filling operation on the first feature vector using a missing value filling module to obtain a second feature vector; the missing value filling module includes learnable parameters for the N feature items, and the filling operation includes encoding the learnable parameter corresponding to the at least one feature item as a corresponding feature value; a processing unit configured to process the second feature vector using a risk prediction model to obtain a risk prediction result for the sample user; An updating unit is configured to update the missing value filling module and the parameters in the risk prediction model according to the risk prediction result and the risk label of the sample user.

12. A device for predicting risk for a user, comprising: a constructing unit configured to construct a first feature vector based on available feature values of a target user for N preset feature items, wherein the target user has a missing feature value for at least one feature item among the N feature items; a filling unit configured to perform a filling operation on the first feature vector using a missing value filling module to obtain a second feature vector; the missing value filling module includes learnable parameters for the N feature items, and the filling operation includes encoding the learnable parameter corresponding to the at least one feature item as a corresponding feature value; The learnable parameters are determined by training the risk prediction model and the missing value filling module; The processing unit is configured to process the second feature vector using a risk prediction model to obtain a risk prediction result of the target user.

13. A computer program product comprising a computer program / instruction, which implements the steps of the method according to any one of claims 1 to 10 when executed by a processor.

14. A computing device comprising a memory and a processor, characterized in that: The memory stores executable code, and when the processor executes the executable code, the method according to any one of claims 1 to 10 is implemented.

Images